TABLE 5.

Comparison between classical classifiers and proposed transfer learning method combined with Cnet1D, when different sizes of target’s training subset are used for training for Nearlab dataset (N = 11) with 8 movement classes (chance accuracy level = 12.5%).

Portion of the training data		KNN (ITD)	LDA (Full)	MLP (ITD)	SVM (ITD)	Cnet1D	PFCnet applied to Cnet1D
1/3	Average accuracy	82.76	86.42	84.90	85.88	86.55	88.43
	Std	6.27	7.11	6.018	6.18	6.18	5.95
	P-value	0 (0.00335)	0 (0.03666)	0 (0.01637)	0 (0.00992)	0 (0.00335)	–
	Alpha (adjusted threshold)	0.0125	0.05	0.025	0.01667	0.01	–
2/3	Average accuracy	84.55	87.74	86.68	86.73	89.68	91.23
	Std	6.21	7.10	5.70	5.92	4.77	4.47
	P-value	0 (0.00335)	0 (0.00335)	0 (0.00444)	0 (0.00333)	0 (0.01279)	–
	Alpha (adjusted threshold)	0.0125	0.01667	0.025	0.01	0.05	–
3/3	Average accuracy	89.20	92.55	91.45	91.72	92.60	93.30
	Std	4.35	4.23	2.97	3.60	3.26	3.41
	P-value	0 (0.00585)	1	0 (0.00764)	0 (0.00334)	0 (0.00992)	–
	Alpha (adjusted threshold)	0.0125	0.05	0.01667	0.01	0.025	–

The p-values are acquired by the pairwise Wilcoxon test when PFCnet applied to Cnet1D is compared with other options. Holm’s method was applied to significance thresholds to calculate alpha.